Nonlinear causal network learning via Granger causality based on extreme support vector regression.

Area of Science:

• Complex Systems Science
• Network Science
• Causal Inference

Background:

• Understanding causality in complex networked systems is challenging due to nonlinearity and time delays.
• Existing methods often struggle to capture the nuanced interactions within these systems.

Purpose of the Study:

• To propose a novel general method for nonlinear causal network learning.
• To accurately identify and quantify causal interactions in complex systems, accounting for time-varying influences.

Main Methods:

• Developed extreme support vector regression Granger causality (ESVRGC), incorporating nonlinearity and nonuniform time-delayed influences.
• Formulated restricted and unrestricted Granger causality models using extreme support vector regression with time-delayed system variables.
• Calculated a nonlinear conditional Granger causality index to measure causal interaction strength.

Main Results:

• ESVRGC demonstrated superior performance compared to popular methods in simulations of nonlinear vector autoregressive and discrete time-delayed dynamic systems.
• Validated the method's robustness across diverse network types, sample sizes, noise levels, and coupling strengths.
• Confirmed the superiority of ESVRGC through experimental studies on real benchmark datasets.

Conclusions:

• ESVRGC provides an effective and robust approach for nonlinear causal network learning.
• The method accurately captures complex causal relationships, including time-delayed effects, in networked systems.
• ESVRGC offers a valuable tool for analyzing real-world complex systems.